From: hal@finney.org
Date: Thu Jun 15 2000 - 12:15:48 MDT
Robin writes, quoting Hal:
> >However upon closer study, I noticed some things which made me more
> >skeptical. His model often lags the actual stock price movements.
> >This is especially noticeable in the 1929 crash. ...
> >But actually you can argue that the drop in dividends in 1930 and
> >onward was caused in part by the stock market crash. ...
> >It seems, therefore, that the stock market crash led to dividend
> >reductions, which then caused his model to predict a stock market crash.
> >But that's not a good prediction because the causative condition only
> >occured after the supposed effect.
>
> The paper I cited (which I just emailed to you) has some statistical
> tests that should not have this problem.
Thanks, I also found the paper online at
http://econ161.berkeley.edu/pdf_files/Fluctuate.pdf. BTW I think the
title is a little joke, based on JP Morgan's famous comment when asked
what the market would do: "it will fluctuate".
This version of the paper was harder for me to follow, apparently part
of an ongoing debate in the research community where I am not familiar
with all the background. But the basic model was essentially the same,
calculating "warranted" (i.e. justifiable by a rational investor) stock
prices based on estimated future dividends, and making those estimates
based on current and past dividend levels. Again the authors are able
to show that much of the variability in stock prices over the longer term
(10-20 years) is captured by this model.
However I feel that my complaint is still valid, that the model could
be fundamentally wrong and still show correlation to actual prices,
by implicitly and subtly including real prices in the calculation.
Real prices affect dividend levels, and so a model which sets warranted
prices based on dividends may be expected to correlate with real prices,
after some time lag. Since real prices correlate with themselves over
the short term the result is that this kind of model can show correlation
even though it has no actual predictive value.
The specific statistic used is the variance of 20 year price changes
versus changes in the modeled price. But I think that if you looked at
20 year price changes against the 20 year price changes lagged a year
or two, there would be a very close relationship and the lagged prices
would "explain" much of the actual variation. Any model which produces
reasonably correct prices after a time lag can look good by this statistic
but is very questionable as a true explanation.
Their figure VI, which shows price/dividend ratios, illustrates that
their model does not predict this value very well. Obviously if you
looked at the deltas, so that you'd have delta-price/delta-dividend
ratios, the model would be even worse, essentially useless. But this
is where I see the fundamental failing of their model being exposed,
as actual prices may lead dividend changes, while their modeled prices
must always follow them. (Of course this kind of data is incredibly
noisy so it is not a very meaningful way to test a model.)
> Another interpretation of the lag is that stocks embody information
> about future dividends that are not embodied in this history of
> past dividends.
That's possible, but it is not what the paper is trying to model. If
this is what is happening then investors must get that information from
somewhere, and it should be possible to come up with an even better model
that uses that source of information and predicts prices more closely.
> >Also, whenever I read these kinds of papers I want to see whether the
> >analysis works when extrapolated forward. ...
> >I doubt that it would predict the degree of increase the market
> >has actually seen, with P/E ratios much higher than historical averages.
>
> My doubt isn't as strong as yours, but yes it would be interested to see.
It's amazing when you look at these charts, which extend to about
1990, and see how they would have to change to hold data up to the
present. Markets are up by something like a factor of 5 in the past
decade, and even counting for inflation, the charts would have to be 50%
taller and go almost straight up to hold the data for the past 10 years.
Hal
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:29:13 MST